TLDR;
- DeepSeek unveils a new version of its R1 model that runs on a single GPU
- Based on Alibaba’s Qwen3-8B, the model excels in complex reasoning tasks
- The distilled version balances performance and efficiency
- Concerns arise over censorship and free speech in the AI’s responses
Chinese artificial intelligence research lab DeepSeek has launched a distilled version of its powerful reasoning model, R1, capable of running efficiently on a single GPU.
The new model, dubbed “DeepSeek-R1-0528-Qwen3-8B,” promises high-level reasoning performance without the need for massive computational infrastructure.
While most high-end AI models require costly, multi-GPU setups, this distilled version is optimized to operate with just one high-performance GPU , such as Nvidia’s H100 , significantly lowering the barrier for developers and researchers to experiment with advanced AI. This is achieved through a method known as knowledge distillation, where a smaller model learns from a larger, more complex one, maintaining core capabilities while improving speed and resource efficiency.
Qwen3 Meets R1
The new release is built upon the recently unveiled Qwen3-8B model by Alibaba, a benchmark in Chinese AI development. By fine-tuning R1 using Qwen3-8B’s architecture, DeepSeek created a hybrid that’s both intelligent and lightweight.
🚀 DeepSeek-R1-0528 is here!
🔹 Improved benchmark performance
🔹 Enhanced front-end capabilities
🔹 Reduced hallucinations
🔹 Supports JSON output & function calling✅ Try it now: https://t.co/IMbTch8Pii
🔌 No change to API usage — docs here: https://t.co/Qf97ASptDD
🔗… pic.twitter.com/kXCGFg9Z5L— DeepSeek (@deepseek_ai) May 29, 2025
According to DeepSeek, the model outperformed Google’s Gemini 2.5 Flash in AIME 2025, a challenging benchmark consisting of advanced mathematical reasoning tasks. Despite its leaner size, the model holds its own in tasks that usually require vast computational power — a rare feat in the AI space.
The model is now publicly available on Hugging Face, the popular platform for open-source AI models, under a permissive license. This opens the door for developers to customize and improve the model for various applications, from education to research and enterprise use.
Speech Limits
Despite the technical triumph, the release has sparked debate over free speech limitations. Early users of the model report that it actively avoids engaging with controversial political topics, particularly those involving the Chinese government.
A developer using the alias xlr8harder criticized the model’s hesitance to directly address human rights issues, such as the internment camps in Xinjiang. “It’s a step backward for open dialogue,” the developer said. Still, they acknowledged the model’s open-source nature could allow the community to overcome these limitations through customization.
Deepseek R1 0528 is substantially less permissive on contentious free speech topics than previous Deepseek releases.
It's unclear if this indicates they've adapted their post-training goals, or if this is another example of a reasoning model. pic.twitter.com/BPOYodBCAH
— xlr8harder (@xlr8harder) May 29, 2025
Notably, this isn’t DeepSeek’s first brush with politically sensitive AI behavior. The company, already under close government scrutiny, has implemented more stringent internal policies in recent months. Some fear these constraints could shape the behavior of its public-facing models in ways that limit their usefulness outside China.
Building Momentum: From Prover to R1
That said, the R1 update comes just a month after DeepSeek quietly released Prover-V2, a specialized AI model focused on formal mathematics and theorem proving. Built on DeepSeek V3 with a staggering 671 billion parameters, Prover-V2 was made public without fanfare, hinting at the lab’s growing confidence in releasing open models.
